Deep learning MPI super-resolution by implicit representation of the system matrix

Abstract

Image reconstruction in MPI is often performed with the system matrix (SM) approach, where the signal of a reference particle sample is measured on a predefined grid in the scanner. This calibration measurement is not only very time-consuming but also places an upper limit to the spatial resolution of the reconstructed image, given by the spacing between two adjacent SM grid points. Recently, implicit neural representations have shown great results in computer vision. They allow for oversampling to gain a higher frequency explicit representation of an object without fixing a certain upsampling scale. We show that this can be used to mostly restore an SM with up to 16x subsampling in 2D and to generate SMs of arbitrary size as an additional tool for image quality improvement. However, we also found that classic spline interpolation is a reasonable tool for this task as well.